Study Before the Ticket Arrives: Pre-Task Preparation for Enterprise Agents

TL;DR for operators

Enterprise agents repeatedly entering the same document stores, applications, and tool environments often spend task-time compute rediscovering where information lives and how the environment behaves. The alternative tested here is to perform some of that discovery before real requests arrive and preserve what was learned as reusable artifacts.

Samuel et al.’s Studying Without a Syllabus: Task-Agnostic Environment Preprocessing¹ reports that 21 of 24 studied method-by-benchmark conditions improve over running the frozen solver with no preparation. The stronger result is not that more preparation is always better. A preparation strategy that works well in one environment can perform poorly in another, and larger study budgets do not reliably increase downstream reward.

For operators, this makes pre-task preparation a systems-design decision rather than simply another inference setting. The relevant questions are what the agent should learn about a stable environment before requests arrive, how that knowledge should be materialized, and whether the one-time cost can be amortized across enough future tasks.

The agent can know the workplace before it knows the work

Consider an enterprise agent connected to the same internal files, applications, and command-line tools every day. The requests may change, but much of the environment does not. Yet a conventional agent can repeatedly spend tokens and tool calls learning folder structure, locating useful documents, discovering application behavior, or reconstructing procedures.

The paper isolates that recurring cost. Before the true downstream tasks are available, a studying system may explore the environment and create reusable files or other contextual resources for a frozen downstream solver. It cannot use real downstream prompts, task traces, labels, verifier outputs, or evaluation feedback.

The authors call this task-agnostic environment preprocessing. The distinction is important: task-agnostic does not mean environment-blind. The studying system can inspect the files and tools it will later operate around; what it cannot inspect is the actual distribution of future tasks.

The downstream solver itself is not fine-tuned. The intervention is the studied environment—the navigation aids, procedures, indices, scripts, corrections, or other artifacts placed around the frozen solver before deployment.

One preparation recipe does not fit every environment

The experiments compare two fixed preparation methods with two more open-ended studying systems.

PREPING generates synthetic practice and distills the resulting trajectories into procedural guidance. Corpus2Skill restructures a document corpus into a navigable skill hierarchy. Each embodies a predetermined idea of what useful preparation looks like.

The Meta-Agent instead decides how to investigate the environment and what artifacts to produce. A second variant receives a workflow archive containing PREPING, Corpus2Skill, and general exploratory-study procedures that it may invoke, combine, or ignore.

Across six heterogeneous benchmarks, a Meta-Agent variant achieves the highest Avg@3 and Best@3 result on five. The archive-assisted Meta-Agent ranks first or second on every benchmark under both metrics.

But the specialist methods reveal why environment structure matters. Corpus2Skill achieves the best Avg@3 score on the very large BCP-G corpus: 0.471, compared with 0.197 for No Study. On OfficeQA, Harvey LAB, and DABStep, however, its Avg@3 score falls below No Study. PREPING is more consistent—it improves over No Study on all six benchmarks—but never ranks first.

This is evidence for adapting the preparation procedure to the environment, not evidence that the Meta-Agent has already solved strategy selection. Archive access itself is uneven. On Avg@3 it helps on four benchmarks and hurts on two, with its largest reported gain, 0.129, on DABStep. On BCP-G, the archive-assisted Meta-Agent does not include Corpus2Skill even though Corpus2Skill is the strongest standalone method there.

More study budget is not a reliable control knob

A natural deployment assumption is that if pre-task study helps, more study compute should help more.

The paper tests that assumption separately with study budgets of $1, $5, $10, $25, and $50 on OfficeQA, Harvey LAB, and Apex Agents. The pattern is not generally monotonic. Sustained scaling appears on Harvey LAB for the Meta-Agent variants, but OfficeQA and Apex Agents remain flat or irregular as the budget rises.

The study also finds that broader exploration of an environment does not reliably predict higher downstream reward.

Those results change how study budget should be interpreted. A larger allowance gives the studying policy more opportunity to act; it does not guarantee that the policy knows how to allocate the additional compute productively. The current experiments therefore support budgeted preparation, but not the rule that increasing the budget is itself a dependable improvement strategy.

Preparation can move compute out of the request path

The other economic question is where the computation occurs.

The paper finds that prepared artifacts can reduce the number of task-time attempts required to reach a given score. Across the six benchmarks, No Study needs between two and eight rollouts to exceed the strongest studied Best@1 score. The estimated inference cost of those additional task-time attempts is 1.6 to 5.5 times the cost of the corresponding studied rollout.

That creates an amortization pathway: spend once to understand a relatively stable environment, then reuse the resulting artifacts across future requests instead of repeatedly rediscovering the same structure.

It does not establish immediate total-cost savings. Upfront study costs vary sharply. For example, average archive-assisted Meta-Agent study cost reaches about $584 on Apex Agents, while PREPING costs about $682 there. The main cross-method performance tables also do not hold study budgets equal, so they should not be read as cost-efficiency rankings.

Cognaptus inference: the strongest use case is therefore not an environment visited once. It is a stable environment serving enough repeated tasks that navigation, procedural, indexing, or verification knowledge has time to repay its construction cost.

Persistent artifacts become part of the agent’s control surface

Reusable preparation also introduces a failure mode that additional task-time sampling does not have in exactly the same form: a persistent artifact can repeatedly steer future searches.

The paper provides both helpful and harmful cases. In one Apex Agents example, an artifact directs the solver toward general pricing material while drawing attention away from files more relevant to the specific task. The study’s artifact-use analysis is descriptive rather than causal, so it does not establish how much of the reward change was caused by artifact use. It does show that preparation can encode an incomplete map of the environment.

For deployment, that means study artifacts should be governed as operational components rather than treated as harmless notes. Provenance, access controls, freshness review, artifact logging, and mechanisms for replacement or rollback become relevant when an artifact can influence many later requests.

What the evidence does—and does not—establish

The evidence is comparatively strong within the tested setup: six heterogeneous benchmarks covering 36 environments, repeated study runs, repeated downstream evaluations, budget sweeps, cost telemetry, and trace-based analyses.

Its external boundary is narrower. The studying agents and downstream solver all use models from the Claude family and share the Claude Code harness. The paper therefore does not establish solver-independent transfer. Only one workflow archive is tested, and there is no non-agentic selector choosing among the same archived workflows, so the experiments cannot isolate how much value comes specifically from agentic strategy selection rather than simply choosing an appropriate fixed method.

The more durable result is architectural. A deployed agent need not treat every incoming request as the first time it has seen its environment. Some computation can happen earlier, produce reusable external knowledge, and change later performance without changing the solver’s weights.

The harder engineering problem is deciding what deserves to be learned in advance. These experiments suggest that this decision cannot be reduced to allocating more study compute. Preparation policy, environment structure, reuse frequency, cost, and artifact governance all enter the equation.

For organizations operating agents repeatedly inside stable specialized environments, that makes the pre-task phase a legitimate design surface—one that can improve performance and reduce repeated task-time search, but only if the preparation itself is treated as something that must be evaluated.

Cognaptus: Automate the Present, Incubate the Future.